ZML's LLMD Wants to Be the Inference Layer That Doesn't Care What Silicon You're Running

The inference stack has a vendor lock-in problem. Most production LLM deployments are quietly coupled to a single GPU architecture — write your serving code for Nvidia, and you've tacitly committed to Nvidia pricing, Nvidia availability, and Nvidia's roadmap. Paris-based startup ZML is making a direct bet that operators are ready to break that coupling.

What LLMD Actually Does

ZML launched LLMD, an LLM inference server engineered to run open-weight models across fundamentally different hardware without rewriting serving logic. The server supports Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc GPUs — not as a future roadmap, but as the declared design surface at launch. The target is data centers and on-premises deployments that already operate mixed accelerator fleets, or that want the freedom to route workloads across whatever silicon is available at a given cost point.

The practical upshot: a single serving interface sits above the hardware layer, abstracting away the differences between GPU vendors and purpose-built accelerators. For operators deploying open-weight community or enterprise models, that means the inference configuration doesn't have to change when the underlying compute does.

The Capital and the Names Behind It

ZML raised $20 million to build and scale LLMD — a round notable less for its size than for who wrote the checks. Backers include Yann LeCun, Meta's chief AI scientist and one of the foundational figures in modern deep learning; Solomon Hykes, the co-founder of Docker, whose containerization work made heterogeneous software deployment tractable in a previous infrastructure era; and several founders from Hugging Face, the platform that has become the de facto distribution hub for open-weight models.

The backer list reads as a deliberate signal. LeCun's presence connects LLMD to open-model advocacy. Hykes brings exactly the infrastructure-abstraction credibility that makes the chip-agnostic pitch legible — Docker solved the "runs on my machine" problem for software; ZML is framing LLMD as the equivalent for inference hardware. The Hugging Face contingent links the product directly to the model ecosystem it intends to serve.

Why Open-Weight Models Change the Infrastructure Calculus

The inference market has historically been shaped by closed model APIs, where the provider owns both the model and the serving stack, eliminating the hardware choice question entirely. Open-weight models break that coupling from the model side — but until the serving layer catches up, operators still end up re-coupling on the hardware side by defaulting to whatever inference stack has the best Nvidia support.

ZML's positioning targets that gap. By centering LLMD on open-weight models and treating multi-vendor hardware as a first-class requirement rather than a stretch goal, it's building for the segment of the market that has already left proprietary model APIs behind but hasn't yet found a serving solution that matches the flexibility they expected.

For data centers running heterogeneous compute — increasingly common as procurement cycles, supply constraints, and cost optimization push operators toward AMD and custom accelerators — a unified inference interface isn't a convenience feature. It's the difference between a coherent deployment architecture and a patchwork of per-chip serving configs.

The Bigger Shift

ZML's launch is a bet on a specific infrastructure thesis: that the accelerator market will remain fragmented, that open-weight models will continue gaining enterprise ground, and that the combination of those two forces will make chip-agnostic inference a durable category rather than a transitional workaround. The $20 million and the credibility of its backers give ZML enough runway to find out whether operators are ready to make that trade. If they are, the inference layer — not the GPU vendor, not the model provider — becomes the durable point of control in the open-model stack.